Group-wise co-salient object detection via multi-view self-labeling novel class discovery

The dominant framework inputs one or two groups of images manually annotated with pixel-level and category labels (i.e. Apple and Horse), and then uses these supervisory signals to train a model in a supervised-learning manner. The trained model is over-fitting to the known categories that cannot generalize well on the novel categories.

To solve the problems, Yang Wu et.al. published their new research on 15 April 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

They proposed a self-labeling NCD framework can effectively transfer the semantic knowledge from the known category (i.e. Apple) to improve clustering of the unknown category, which can generalize well to the novel category.

In this paper, they have presented a self-labeling NCD frame-work for CoSOD. Among it, a multi-view self-labeling strategy via the SK algorithm has been presented to effectively transfer the semantic knowledge of the known classes to improve the clustering of the unknown classes, which helps improve the generalization capability of the model to novel categories. Besides, they have designed two effective modules of DRFM and GIM, among which the DRFM is to extract the locally compact features, while the GIM is to maximize inter-group separability. Extensive experiments on Cosal2015, CoSOD3k, CoCA have demonstrated that our method has achieved superior performance compared to the state-of-the-arts in terms of all evaluation metrics with a fast speed of 50 fps.

DOI: 10.1007/s11704-023-3284-5